Fast Fluid Detection in 3-D Macular OCT Scans

Abstract

 

Purpose

 

A novel tool to detect intraretinal cysts and subretinal fluid in OCT C-scans is presented. Its purpose is to speed-up and objectifies OCT examinations by telling the ophthalmologist whether or not there is fluid in the retina and, if there is, to indicate in an en-face localizer where the fluid lies.

 

Methods

 

One retina expert manually delineated every fluid pocket in 30 macular C-scans from 30 AMD patients. Ten of these scans did not contain any fluid. Additionally, 30 macular scans from healthy subjects were also included in this study. Each C-scans was acquired by a Heidelberg Spectralis OCT system. The ILM and the RPE were jointly segmented in subsampled C-scans using a state-of-the-art graph-based algorithm from our group. Then, the intensity was normalized in each A-scan in order to highlight the fluid pockets. This normalization relies on a statistical model of the intensity decrease from the ILM to the RPE across the macula. Finally, a support vector machine was trained by two-fold cross-validation to map the intensity distribution in each A-scan to a local fluid probability.

 

Results

 

By varying a threshold on the local fluid probability, the algorithm was able to localize fluid with an area under the ROC curve of 0.953. As for the binary “fluid / no fluid” decision, it was able to classify C-scans with a specificity of 90% and a sensitivity of 100%. The average processing time was 13 seconds per C-scan. This includes eight seconds for layer segmentation, which may already be available from another task.

 

Conclusions

 

We have presented a tool able to inform the ophthalmologist in seconds whether or not he or she needs to look for fluid in a macular C-scan. The tool is quite specific: a false alarm rate of 10% was achieved without missing any pathological cases. Additionally, a 2-D fluid probability map, which could be overlaid on the scanning laser ophthalmoscopy localizer image, was defined to quickly navigate to the relevant regions of the retina and therefore speed-up the OCT examination further.

 

 

Segmentation pipeline. Figures A and B show one sagittal and one coronal view of a subsampled C-scan. Figure C shows the intensity normalized C-scan between the ILM and the RPE. Figure D show the en-face fluid probability map.

 

Keywords: 412 age-related macular degeneration • 550 imaging/image analysis: clinical • 496 detection